David Fields

Nothing to see here… Just Jensen Huang (CEO of the world’s most valuable company Nvidia) and Chamath discussing Bittensor $TAO 🤯

Bittensor will be run by agents. They will feed the mining, resist the exploits, manage the fleets, build the subnets and consume the commodities

We just completed the largest decentralised LLM pre-training run in history: Covenant-72B. Permissionless, on Bittensor subnet 3. 72B parameters. ~1.1T tokens. Commodity internet. No centralized cluster. No whitelist. Anyone with GPUs could join or leave freely. 1/n

👀 something new is coming We've been building and we're almost ready to show you. SN33 has been processing the web at scale, turning raw Common Crawl data into clean, AI-ready `llms.txt` files. Structured semantic summaries that any LLM agent, MCP server, or AI app can consume instantly. On Thursday we'll be releasing the Github repo where `llms.txt` files will be pushed in batches as the subnet processes them. We're starting with over 1000 websites analyzed and processed by the subnet that will grow every week. And shortly after... 🌍 We're launching a public frontend Any website. Any domain. You request it, the subnet processes it and you get a `llms.txt` back. No more raw HTML hell for AI agents. No more redundant crawling. Just clean, structured, machine-readable intelligence about any corner of the web, on demand, powered by decentralized compute. This is SN33 becoming a public utility for AI infrastructure The web, made readable for machines. At scale. Open to anyone. 🔜 More very soon. Stay tuned.

SN33 -- Enriching the Data of the World SN33 just shipped Webpage Metadata v2, and the best way to explain what we’re building is this: an llms.txt version of Common Crawl. Our partnership with Common Crawl began with the simple but daunting task of tagging web pages to make semantic web data widely available. Generating this data would break down the barriers preventing web organization. This week we are taking a giant step in broadening that goal to encapsulate AI-enabling the world wide web by launching the enrichment process for entire web sites. Search engines atomize the web by surfacing individual pages. That's great for finding individual facts. It does almost nothing to give agents the holistic information they need to actually complete tasks. Simple example: an agent searching for "best skis" gets quality-for-price rankings from individual pages. It completely misses how waist width affects your ability to float in powder, navigate tight spaces, or carve on groomed trails. That information exists across an entire site, but no one is structuring it that way. This week we shipped the technology to change that. SN33 is now enriching entire web sites, not just individual pages. Our new high-volume API pushes full sites through the subnet, collecting enriched data from tags, NER, similar pages to summarization across every page on a site, grouped together. Why llms.txt matters The llms.txt standard summarizes an entire web site's contents in a single meaningful text file. Agents and MCP tools can understand what a site contains without processing every page. It's the missing layer between the open web and the agent economy. Adoption has been stymied by one problem: nobody is generating these files at scale. There hasn't been a broad effort to create llms.txt for the whole web — until now. Once SN33 reaches tipping-point volume of enriched site data, we begin publishing llms.txt files at scale. We believe SN33 will become the largest producer of llms.txt files in the world. The demand for structured web data is already proven. Our first open-source dataset, the 5000 Podcast Conversations, has crossed 300,000+ downloads on HuggingFace. That was conversations. This is the entire open web. More open-source releases are coming. v2.28.63 is on testnet and goes mainnet February 23rd.

update just earned my first $1 on subnet 33. up only from here.

SN33 🤝 Chutes Integration Chutes is now a the preferred inference provider in our LLM backend. Subnets powering subnets: → SN33 miners/validators can route inference through Chutes' decentralized compute → Volume and utility flow within the ecosystem, not to OpenAI → Reduces single points of failure in validation pipeline Also adding OpenRouter for model optionality and redundancy. The Bittensor flywheel works best with subnets building on each other instead of routing value to Web2 infrastructure. Testnet live. Mainnet next week. PR and more detail on v2.26.61 in Discord.

New Revenue Generating Partnerships for ReadyAI We're launching Named Entity Recognition (NER) as our next task type as we continue to enable the full universe of structured data functionality on the subnet. NER is one of the largest use cases for structured data in AI — extracting and classifying key information from unstructured text at scale. It's foundational infrastructure that powers everything from search to compliance to market intelligence. We're launching it with a real estate focused use case: Regulatory Radar. We process regulatory information from city council meetings, zoning decisions, permit approvals, and municipal records — extracting the entities, relationships, and signals that move real estate markets. Information that used to take teams hours to find is now structured and actionable. This is part of our AcquiOS platform, and we already have 2 enterprise customers using it: Gelt Venture Partners and Archer Equities — with a strong pipeline of additional RE firms behind them. Massive task type. Real customers using it already. More to come. Onward!